# FEEDBACK GUARD 0.1.0 manual

## What it does

Original 24-slot causal notch processor for mono/stereo VST3 buses. Use on a monitor send, master bus or individual channel.24 stable slots share Manual, Fixed and Live duties: an enabled manual slot reserves its place; automatic Fixed/Live allocations use remaining slots within their limits. No automatic boost, frequency shifter, hidden limiter or room correction sweep is included.

All 192 physical controls are host automatable. The wrapper uses the last valid automation point per processing block; continuous controls transition smoothly. It does not claim sample-accurate automation. Learn, Stop, Clear and Freeze are deliberate queued UI actions, not host automation parameters.16 factory presets, A/B comparison and learned banks are saved in versioned state.

## Modes and workflow

- **Manual:** enable individual slots and adjust their cuts; retained learned protection remains available.
- **Learn Fixed:** select this mode, then deliberately press Start Learn. Detection captures Fixed notches. Stop Learn stops discovery and retains these filters. Changing mode or loading a preset never starts learning by itself.
- **Live Assist:** persistent narrow-band evidence can allocate adaptive Live filters. They hold and recover when evidence disappears.
- **Freeze Live:** converts current Live filters into retained Fixed filters.
- **Clear Fixed / Clear Live:** separately remove the chosen learned bank types. Output-active destructive clearing and bypass have confirmation in the editor. Manual filters are controlled by their own Enable values.

Linked Stereo detects a common spectrum and applies common cuts. Dual Mono detects and allocates independently to the stimulated left/right side. Both share the same 24-slot capacity. On a mono host bus, the mono input is duplicated internally and summed back to mono.

Presets change controls while preserving learned protection; clearing is explicit. A/B stores both controls and learned-bank state. Copy copies the current comparison snapshot; Swap exchanges the two. Host project saves include pending valid learned-bank restores and deliberate clears, so saving need not wait for an audio callback. Transient detection history is not saved. Restored Live filters start with a fresh hold time.

## Global controls

| Control | Purpose |
|---|---|
| Mode / Routing | Manual, Learn Fixed, Live Assist; Linked Stereo or Dual Mono. |
| Input / Output | Input trim−24…+12dB and final output trim−24…+12dB. |
| Mix |0…100% dry/wet blend. Output trim still applies to the blend. |
| Bypass | Product bypass transitions to exact untrimmed dry; standard host bypass is also supported. |
| Sensitivity |0…100%, adjusts the narrow spectral prominence requirement. Higher can cause more musical false positives. |
| Noise Floor |−90…−20dBFS, rejects candidates below the detection floor. |
| Min / Max Frequency | Bounds automatic detection. Manual notch frequencies can cover20Hz…20kHz. |
| Persistence |100…2000ms of qualifying evidence before allocation. |
| Prominence |3…30dB required spectral prominence over surrounding bins. |
| Growth |0…24dB/s growth evidence.0 deliberately allows stationary isolated tones. |
| Harmonic Guard |0…100%, rejects candidates with harmonic context more strongly as increased. This is a heuristic. |
| Auto Depth / Auto Q | Automatic cut3…24dB and Q2…60. |
| Fixed Limit / Live Limit |0…24 per type, limited by remaining shared slots.0 prevents allocations of that type. |
| Live Hold / Recovery | Hold0.1…60s, then recover over0.1…30s if the resonance remains absent. |
| Attack |1…200ms cut transition. |
| Cooldown |0.1…10s between new automatic allocations. |
| Dynamic Knee |0…12dB transition around manual dynamic threshold. |
| Analyzer | Enables visible input/output spectra; required detector analysis still runs when needed. |

## Each slot

| Control | Purpose |
|---|---|
| Enabled | Reserves this slot for a manual notch. |
| Frequency |20Hz…20kHz centre, bounded below Nyquist for the current sample rate. |
| Depth |0…36dB attenuation; no boosts. |
| Q |0.5…80; higher means a narrower cut. |
| Dynamic | Enables signal-dependent manual attenuation. |
| Threshold |−80…0dBFS detector threshold for the dynamic cut. |
| Release |20…5000ms dynamic recovery. |

The slot bank shows effective frequencies, attenuation, confidence and Fixed/Live/Manual state. Graph selection/drag changes frequency horizontally and cut depth vertically; the mouse wheel changes Q/bandwidth for an enabled manual slot. Shift provides finer adjustments. Controls and typed values expose full precision. Input/output FFT and filter response are calculated from actual signal/filter state. The response uses a coherent snapshot of applied filter coefficients and blends, with per-pixel sampling plus exact active notch centres and nearby bandwidth points, so narrow cuts remain visible. The history represents real filter activity. Meter levels refer to the digital stream, not calibrated acoustic SPL.

## Detection time and audio latency

The audio path uses causal IIR notches and declares **0 samples PDC**, at supported rates22.05…192kHz. Analysis uses an 8192-sample window and 512-sample hop, plus bounded incremental FFT scheduling finishing within 495 samples after its window ends. At 48kHz the window is about 171ms and the additional scheduling about 10.3ms, before persistence/qualification. These are detection response times; they do not delay the causal audio stream. Hardware/host buffers still add end-to-end latency.

The conservative host-reported IIR tail is 22 seconds for retained filters, including extreme narrow low-frequency banks. This allows the host to preserve filter decay and is not a room reverb or prediction of an acoustic feedback loop.

## Practical limits and safe operation

Feedback detection uses persistence, spectral prominence, amplitude/growth, harmonic context, frequency limits and cooldown. A sustained musical sine and a feedback sine can be identical at this input; the processor cannot always distinguish them. Automatic cuts may alter a voice/instrument or miss feedback. More sensitive/faster settings increase that tradeoff. Do not advertise or expect universal protection, gain-before-feedback improvement in every room or instant response.

This release is physically unqualified for microphone/speaker ringout. Numerical simulations and actual DLL tests are useful engineering evidence, not proof of performance in your room. Begin at low monitored system levels, maintain direct mute/gain control and evaluate in supervised conditions. No automated loud ringout, test sweep or gain-raising procedure is provided. Room measurement/correction is a separate task.

## Validation and troubleshooting

Numerical tests include independently measured notch attenuation/Q, fixed/live retention/recovery, all 192 controls, all 16 silent presets, linked/dual-mono behaviour, extreme values, silence, bypass, five rates and block partitioning. Actual DLL SDK tests cover mono/stereo float32/64, automation/state, malformed state rejection, reference audio and zero PDC. Pluginval strictness5 and10 passed. Finite local CPU tests include 24 manual filters with analyzer enabled; OS/driver stalls and other computers can differ. No claim of immunity from crashes or dropouts is made.

If there are no automatic cuts, check mode, input level, frequency limits, persistence, Growth, available slots and limits. Start Learn is needed for Fixed discovery. If music is affected, reduce sensitivity/depth or strengthen persistence/harmonic safeguards, and inspect actual filter activity. If manual cuts seem inactive, check Enable, Mix, depth, bypass and the host's selected bus. For dropouts, raise your interface buffer and reduce unrelated processing; the standalone's DirectSound buffer is independent of plugin PDC.

Website https://normsonic.com/ — support info@normsonic.com. Free binary use is governed by LICENSE.txt; your own processed audio remains yours.